“Scaling laws exist for the pre-training phase of AI, showing that model performance predictably improves with increases in compute, dataset size, and neural network size.”